Not sure if this is more of a question for graph models or small molecules --
I'm looking into models for encoding the 3D point cloud, with atom types, of a molecule, to a fixed size vector embedding. I don't care about the adjacency matrix or graph-level features, just the identities & locations of the input nodes.
I've done this in the past with a PointNet encoder and 3D convolutional decoder, but it was slow, heavy to train, and frankly inelegant. The encoder seems straightforward - some equivariant model or PointNet. I have tried a few graph-based decoders with various loss functions, and poked around the graph generation and point cloud literatures, but haven't been able to find a solution that really "clicks".
Asking here to see if there's anything known to work, or something obvious I've missed.